It's a simple program. test environment: debian 8, go 1.4.2
union.go:
package main
import "fmt"
type A struct {
t int32
u int64
}
func test() (total int64) {
a := [...]A{{1, 100}, {2, 3}}
for i := 0; i < 5000000000; i++ {
p := &a[i%2]
total += p.u
}
return
}
func main() {
total := test()
fmt.Println(total)
}
union.c:
#include <stdio.h>
struct A {
int t;
long u;
};
long test()
{
struct A a[2];
a[0].t = 1;
a[0].u = 100;
a[1].t = 2;
a[1].u = 3;
long total = 0;
long i;
for (i = 0; i < 5000000000; i++) {
struct A* p = &a[i % 2];
total += p->u;
}
return total;
}
int main()
{
long total = test();
printf("%ld
", total);
}
result compare:
go:
257500000000
real 0m9.167s
user 0m9.196s
sys 0m0.012s
C:
257500000000
real 0m3.585s
user 0m3.560s
sys 0m0.008s
It seems that the go compiles lot of weird assembly codes (you could use objdump -D to check it).
For example, why movabs $0x12a05f200,%rbp appears twice?
400c60: 31 c0 xor %eax,%eax
400c62: 48 bd 00 f2 05 2a 01 movabs $0x12a05f200,%rbp
400c69: 00 00 00
400c6c: 48 39 e8 cmp %rbp,%rax
400c6f: 7d 46 jge 400cb7 <main.test+0xb7>
400c71: 48 89 c1 mov %rax,%rcx
400c74: 48 c1 f9 3f sar $0x3f,%rcx
400c78: 48 89 c3 mov %rax,%rbx
400c7b: 48 29 cb sub %rcx,%rbx
400c7e: 48 83 e3 01 and $0x1,%rbx
400c82: 48 01 cb add %rcx,%rbx
400c85: 48 8d 2c 24 lea (%rsp),%rbp
400c89: 48 83 fb 02 cmp $0x2,%rbx
400c8d: 73 2d jae 400cbc <main.test+0xbc>
400c8f: 48 6b db 10 imul $0x10,%rbx,%rbx
400c93: 48 01 dd add %rbx,%rbp
400c96: 48 8b 5d 08 mov 0x8(%rbp),%rbx
400c9a: 48 01 f3 add %rsi,%rbx
400c9d: 48 89 de mov %rbx,%rsi
400ca0: 48 89 5c 24 28 mov %rbx,0x28(%rsp)
400ca5: 48 ff c0 inc %rax
400ca8: 48 bd 00 f2 05 2a 01 movabs $0x12a05f200,%rbp
400caf: 00 00 00
400cb2: 48 39 e8 cmp %rbp,%rax
400cb5: 7c ba jl 400c71 <main.test+0x71>
400cb7: 48 83 c4 20 add $0x20,%rsp
400cbb: c3 retq
400cbc: e8 6f e0 00 00 callq 40ed30 <runtime.panicindex>
400cc1: 0f 0b ud2
...
while the C assembly is more clean:
0000000000400570 <test>:
400570: 48 c7 44 24 e0 64 00 movq $0x64,-0x20(%rsp)
400577: 00 00
400579: 48 c7 44 24 f0 03 00 movq $0x3,-0x10(%rsp)
400580: 00 00
400582: b9 64 00 00 00 mov $0x64,%ecx
400587: 31 d2 xor %edx,%edx
400589: 31 c0 xor %eax,%eax
40058b: 48 be 00 f2 05 2a 01 movabs $0x12a05f200,%rsi
400592: 00 00 00
400595: eb 18 jmp 4005af <test+0x3f>
400597: 66 0f 1f 84 00 00 00 nopw 0x0(%rax,%rax,1)
40059e: 00 00
4005a0: 48 89 d1 mov %rdx,%rcx
4005a3: 83 e1 01 and $0x1,%ecx
4005a6: 48 c1 e1 04 shl $0x4,%rcx
4005aa: 48 8b 4c 0c e0 mov -0x20(%rsp,%rcx,1),%rcx
4005af: 48 83 c2 01 add $0x1,%rdx
4005b3: 48 01 c8 add %rcx,%rax
4005b6: 48 39 f2 cmp %rsi,%rdx
4005b9: 75 e5 jne 4005a0 <test+0x30>
4005bb: f3 c3 repz retq
4005bd: 0f 1f 00 nopl (%rax)
Could somebody explain it? Thanks!
The main difference is the the array bounds checking. In the disassembly dump for the Go program, there is:
400c89: 48 83 fb 02 cmp $0x2,%rbx
400c8d: 73 2d jae 400cbc <main.test+0xbc>
...
400cbc: e8 6f e0 00 00 callq 40ed30 <runtime.panicindex>
400cc1: 0f 0b ud2
So if %rbx is greater than or equal to 2, then it jumps down to a call to runtime.panicindex. Given you're working with an array of size 2, that is clearly the bounds check. You could make the argument that the compiler should be smart enough to skip the bounds check in this particular case where the range of the index can be determined statically, but it seems that it isn't smart enough to do so yet.
While you're seeing a noticeable performance difference for this micro-benchmark, it might be worth considering whether this is actually representative of your actual code. If you're doing other stuff in your loop, the difference is likely to be less noticeable.
And while bounds checking does have a cost, in many cases it is better than the alternative of the program continuing with undefined behaviour.